Image transforms¶

vision.transform provides classes and functions to handle data augmentation in computer vision.

Data augmentation¶

If you want to quickly get a set of random transforms that have proved to work well in a wide range of tasks, you should use the get_transforms function. The most important parameters to adjust are do_flip and flip_vert, depending on the type of images you have.

This function returns a tuple of two list of transforms, one for the training set and the other for the validation set (which is limited to a center crop by default.

tfms = get_transforms(); len(tfms)

2

Here is the example image we will use to show the data augmentation.

def get_ex(): return open_image('imgs/cat_example.jpg')
get_ex().show()

Let's see how the defaults of get_transforms change this little kitten now.

tfms = get_transforms()
fig, axs = plt.subplots(2,4,figsize=(12,6))
for ax in axs.flatten():
    img = apply_tfms(tfms[0], open_image('imgs/cat_example.jpg'), size=224)
    img.show(ax=ax)

Another useful function that gives basic transforms is:

scale should be a given float if do_rand is false, otherwise it can be a range of floats (and the zoom will have a random value inbetween). Again, here is a sense of what this can give us.

tfms = zoom_crop(scale=(1.,1.2), do_rand=True)
fig, axs = plt.subplots(2,4,figsize=(12,6))
for ax in axs.flatten():
    img = apply_tfms(tfms[0], get_ex(), size=224)
    img.show(ax=ax)

show_doc(rand_resize_crop, ignore_warn=True, arg_comments={
    'size': 'Final size of the image',
    'max_scale': 'Zooms the image to a random scale up to this',
    'ratios': 'Range of ratios in which a new one will be randomly picked'
})

This transforms determines a new width and height of the image after the random scale and squish to the new ratio are applied. Those are switched with probabilit 0.5, then we return the part of the image with the width and height computed centered in row_pct, col_pct if width and height are both less than the corresponding size of the image, otherwise we try again with new ranfom parameters.

tfm = rand_resize_crop(224)
_, axs = plt.subplots(2,4,figsize=(12,6))
for ax in axs.flatten():
    img = apply_tfms(tfm, get_ex(), size=224)
    img.show(ax=ax)

Randomness¶

The functions that define each transform, like rotateor flip_lr are deterministic. The fastai library will then randomize them in two different ways:

each transform can be defined with an argument named p representing the probability for it to be applied
each argument that is type-annoted with a random function (like uniform or rand_int) can be replaced by a tuple of arguments accepted by this function, and on each call of the transform, the argument that is passed inside the function will be picked randomly using that random function.

If we look at the function rotate for instance, we see it had an argument degrees that is type-annotated as uniform.

First level of randomness: We can define a transform using rotate with degrees fixed to a value, but by passing an argument p. The rotation will then be executed with a probability of p but always with the same value of degrees.

tfm = [rotate(degrees=30, p=0.5)]
fig, axs = plt.subplots(1,5,figsize=(12,4))
for ax in axs:
    img = apply_tfms(tfm, get_ex())
    title = 'Done' if tfm[0].do_run else 'Not done'
    img.show(ax=ax, title=title)

Second level of randomness: We can define a transform using rotate with degrees defined as a range, without an argument p. The rotation will then always be executed with a random value picked uniformly between the two floats we put in degrees.

tfm = [rotate(degrees=(-30,30))]
fig, axs = plt.subplots(1,5,figsize=(12,4))
for ax in axs:
    img = apply_tfms(tfm, get_ex())
    title = f"deg={tfm[0].resolved['degrees']:.1f}"
    img.show(ax=ax, title=title)

All combined: We can define a transform using rotate with degrees defined as a range, and an argument p. The rotation will then always be executed with a probability p and a random value picked uniformly between the two floats we put in degrees.

tfm = [rotate(degrees=(-30,30), p=0.75)]
fig, axs = plt.subplots(1,5,figsize=(12,4))
for ax in axs:
    img = apply_tfms(tfm, get_ex())
    title = f"Done, deg={tfm[0].resolved['degrees']:.1f}" if tfm[0].do_run else f'Not done'
    img.show(ax=ax, title=title)

List of transforms¶

Here is the list of all the deterministic functions on which the transforms are built. As explained before, each of those can have a probability p of being executed, and any time an argument is type-annotated with a random function, it's possible to randomize it via that function.

show_doc(brightness)

This transform adjusts the brightness of the image depending of the value in change. A change of 0 will transform the image in black and a change of 1 will transform the image to white. 0.5 doesn't do anything.

fig, axs = plt.subplots(1,5,figsize=(12,4))
for change, ax in zip(np.linspace(0.1,0.9,5), axs):
    brightness(get_ex(), change).show(ax=ax, title=f'change={change:.1f}')

This adjusts the contrast depending of the value in scale. A scale of 0 will transform the image in grey and a very high scale will transform the picture in super-contrast. 1. doesn't do anything.

fig, axs = plt.subplots(1,5,figsize=(12,4))
for scale, ax in zip(np.exp(np.linspace(log(0.5),log(2),5)), axs):
    contrast(get_ex(), scale).show(ax=ax, title=f'scale={scale:.2f}')

This transform takes a crop of the image to return one of the given size. The position is given by (col_pct, row_pct), with col_pct and row_pct being normalized between 0. and 1.

fig, axs = plt.subplots(1,5,figsize=(12,4))
for center, ax in zip([[0.,0.], [0.,1.],[0.5,0.5],[1.,0.], [1.,1.]], axs):
    crop(get_ex(), 300, *center).show(ax=ax, title=f'center=({center[0]}, {center[1]})')

This works like crop but if the target size is bigger than the size of the image (on one or the other dimension), padding is applied according to padding_mode (see pad for an example of all the options) and the position of center is ignored on that dimension.

fig, axs = plt.subplots(1,5,figsize=(12,4))
for size, ax in zip(np.linspace(200,600,5), axs):
    crop_pad(get_ex(), int(size), 'zeros', 0.,0.).show(ax=ax, title=f'size = {int(size)}')

This transform applies one of all the transformations possible of the image by combining a flip (horizontal or vertical) and a rotation of a multiple of 90 degrees.

fig, axs = plt.subplots(2,4,figsize=(12,8))
for k, ax in enumerate(axs.flatten()):
    dihedral(get_ex(), k).show(ax=ax, title=f'k={k}')
plt.tight_layout()

This transform horizontally flips the image.

fig, axs = plt.subplots(1,2,figsize=(6,4))
get_ex().show(ax=axs[0], title=f'no flip')
flip_lr(get_ex()).show(ax=axs[1], title=f'flip')

This transform changes the pixels of the image by randomly replacing them with pixels from the neighborhood (how far we go is controlled by the value of magnitude).

fig, axs = plt.subplots(1,5,figsize=(12,4))
for magnitude, ax in zip(np.linspace(-0.05,0.05,5), axs):
    tfm = jitter(magnitude=magnitude)
    get_ex().jitter(magnitude).show(ax=ax, title=f'magnitude={magnitude:.2f}')

Pads the image by adding padding pixel on each side of the picture accordin to mode:

mode = zeros pads with zeros,
mode = border repeats the pixels at the border.
mode = reflection pads by taking the pixels symmetric to the border.

fig, axs = plt.subplots(1,3,figsize=(12,4))
for mode, ax in zip(['zeros', 'border', 'reflection'], axs):
    pad(get_ex(), 50, mode).show(ax=ax, title=f'mode={mode}')

Perspective wrapping is a deformation of the image as it was seen in a different plane of the 3D-plane. The new plane is determined by telling where we want each of the four corners of the image (from -1 to 1, -1 being left/top, 1 being right/bottom).

fig, axs = plt.subplots(2,4,figsize=(12,8))
for i, ax in enumerate(axs.flatten()):
    magnitudes = torch.tensor(np.zeros(8))
    magnitudes[i] = 0.5
    perspective_warp(get_ex(), magnitudes).show(ax=ax, title=f'coord {i}')

fig, axs = plt.subplots(1,5,figsize=(12,4))
for deg, ax in zip(np.linspace(-60,60,5), axs):
    get_ex().rotate(degrees=deg).show(ax=ax, title=f'degrees={deg}')

fig, axs = plt.subplots(2,4,figsize=(12,8))
for i, ax in enumerate(axs.flatten()):
    get_ex().skew(i, 0.2).show(ax=ax, title=f'direction={i}')

fig, axs = plt.subplots(1,5,figsize=(12,4))
for scale, ax in zip(np.linspace(0.66,1.33,5), axs):
    get_ex().squish(scale=scale).show(ax=ax, title=f'scale={scale:.2f}')

Apply the four tilts at the same time, each with a strength given in the vector magnitude. See tilt just below for the effect of each individual tilt.

tfm = symmetric_warp(magnitude=(-0.2,0.2))
_, axs = plt.subplots(2,4,figsize=(12,6))
for ax in axs.flatten():
    img = apply_tfms(tfm, get_ex(), padding_mode='zeros')
    img.show(ax=ax)

Tilts c in the direction given (0: left, 1: right, 2: top, 3: bottom) with a certain magnitude. A positive number is a tilt forward (toward the person looking at the picture), a negative number a tilt backward.

fig, axs = plt.subplots(2,4,figsize=(12,8))
for i in range(4):
    get_ex().tilt(i, 0.4).show(ax=axs[0,i], title=f'direction={i}, fwd')
    get_ex().tilt(i, -0.4).show(ax=axs[1,i], title=f'direction={i}, bwd')

fig, axs = plt.subplots(1,5,figsize=(12,4))
for scale, ax in zip(np.linspace(1., 1.5,5), axs):
    get_ex().squish(scale=scale).show(ax=ax, title=f'scale={scale:.2f}')

Last random functions¶

tfm = rand_crop()
_, axs = plt.subplots(2,4,figsize=(12,6))
for ax in axs.flatten():
    img = apply_tfms(tfm, get_ex(), size=224)
    img.show(ax=ax)

tfm = rand_zoom(scale=(1.,1.5))
_, axs = plt.subplots(2,4,figsize=(12,6))
for ax in axs.flatten():
    img = apply_tfms(tfm, get_ex())
    img.show(ax=ax)

Image transforms¶

Data augmentation¶

`get_transforms`

`zoom_crop`

`rand_resize_crop`

Randomness¶

List of transforms¶

`brightness`

`contrast`

`crop`

`crop_pad`

`dihedral`

`flip_lr`

`jitter`

`pad`

`perspective_warp`

`rotate`

`skew`

`squish`

`symmetric_warp`

`tilt`

`zoom`

Last random functions¶

`rand_crop`

`rand_zoom`